maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/X86/fp-intrinsics.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1993 lines
64 KiB
LLVM
Raw Normal View History

[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; RUN: llc -O3 -mtriple=i686-pc-linux -mattr=+cmov < %s | FileCheck %s --check-prefix=COMMON --check-prefix=X87
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; RUN: llc -O3 -mtriple=i686-pc-linux -mattr=sse2 < %s | FileCheck %s --check-prefix=COMMON --check-prefix=X86-SSE
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; RUN: llc -O3 -mtriple=x86_64-pc-linux < %s | FileCheck %s --check-prefix=COMMON --check-prefix=SSE
; RUN: llc -O3 -mtriple=x86_64-pc-linux -mattr=+avx < %s | FileCheck %s --check-prefix=COMMON --check-prefix=AVX --check-prefix=AVX1
; RUN: llc -O3 -mtriple=x86_64-pc-linux -mattr=+avx512f < %s | FileCheck %s --check-prefix=COMMON --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512F
; RUN: llc -O3 -mtriple=x86_64-pc-linux -mattr=+avx512dq < %s | FileCheck %s --check-prefix=COMMON --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512DQ
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
; Verify that constants aren't folded to inexact results when the rounding mode
; is unknown.
;
; double f1() {
; // Because 0.1 cannot be represented exactly, this shouldn't be folded.
; return 1.0/10.0;
; }
;
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f1() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f1:
; X87: # %bb.0: # %entry
; X87-NEXT: fld1
; X87-NEXT: fdivs {{\.LCPI.*}}
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f1:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: divsd {{\.LCPI.*}}, %xmm0
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: fldl (%esp)
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f1:
; SSE: # %bb.0: # %entry
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: divsd {{.*}}(%rip), %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: f1:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vdivsd {{.*}}(%rip), %xmm0, %xmm0
; AVX-NEXT: retq
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
entry:
%div = call double @llvm.experimental.constrained.fdiv.f64(
double 1.000000e+00,
double 1.000000e+01,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
ret double %div
}
; Verify that 'a - 0' isn't simplified to 'a' when the rounding mode is unknown.
;
; double f2(double a) {
; // Because the result of '0 - 0' is negative zero if rounding mode is
; // downward, this shouldn't be simplified.
; return a - 0;
; }
;
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f2(double %a) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f2:
; X87: # %bb.0: # %entry
; X87-NEXT: fldz
; X87-NEXT: fsubrl {{[0-9]+}}(%esp)
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f2:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: xorpd %xmm1, %xmm1
; X86-SSE-NEXT: subsd %xmm1, %xmm0
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: fldl (%esp)
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f2:
; SSE: # %bb.0: # %entry
; SSE-NEXT: xorpd %xmm1, %xmm1
; SSE-NEXT: subsd %xmm1, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: f2:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vxorpd %xmm1, %xmm1, %xmm1
; AVX-NEXT: vsubsd %xmm1, %xmm0, %xmm0
; AVX-NEXT: retq
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
entry:
[FPEnv] Expand constrained FP POWI Modify ExpandStrictFPOp(...) to handle nodes that have scalar operands. Also, add a Strict FMA test and do some other light cleanup in the Strict FP code. Differential Revision: https://reviews.llvm.org/D48149 llvm-svn: 334863
2018-06-16 04:57:55 +08:00
%sub = call double @llvm.experimental.constrained.fsub.f64(
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
double %a,
double 0.000000e+00,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
[FPEnv] Expand constrained FP POWI Modify ExpandStrictFPOp(...) to handle nodes that have scalar operands. Also, add a Strict FMA test and do some other light cleanup in the Strict FP code. Differential Revision: https://reviews.llvm.org/D48149 llvm-svn: 334863
2018-06-16 04:57:55 +08:00
ret double %sub
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
}
; Verify that '-((-a)*b)' isn't simplified to 'a*b' when the rounding mode is
; unknown.
;
; double f3(double a, double b) {
; // Because the intermediate value involved in this calculation may require
; // rounding, this shouldn't be simplified.
; return -((-a)*b);
; }
;
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f3(double %a, double %b) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f3:
; X87: # %bb.0: # %entry
; X87-NEXT: fldz
; X87-NEXT: fchs
; X87-NEXT: fld %st(0)
; X87-NEXT: fsubl {{[0-9]+}}(%esp)
; X87-NEXT: fmull {{[0-9]+}}(%esp)
; X87-NEXT: fsubrp %st, %st(1)
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f3:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movapd %xmm0, %xmm1
; X86-SSE-NEXT: subsd {{[0-9]+}}(%esp), %xmm1
; X86-SSE-NEXT: mulsd {{[0-9]+}}(%esp), %xmm1
; X86-SSE-NEXT: subsd %xmm1, %xmm0
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: fldl (%esp)
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f3:
; SSE: # %bb.0: # %entry
; SSE-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero
; SSE-NEXT: movapd %xmm2, %xmm3
; SSE-NEXT: subsd %xmm0, %xmm3
; SSE-NEXT: mulsd %xmm1, %xmm3
; SSE-NEXT: subsd %xmm3, %xmm2
; SSE-NEXT: movapd %xmm2, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: f3:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero
; AVX-NEXT: vsubsd %xmm0, %xmm2, %xmm0
; AVX-NEXT: vmulsd %xmm1, %xmm0, %xmm0
; AVX-NEXT: vsubsd %xmm0, %xmm2, %xmm0
; AVX-NEXT: retq
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
entry:
%sub = call double @llvm.experimental.constrained.fsub.f64(
double -0.000000e+00, double %a,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
%mul = call double @llvm.experimental.constrained.fmul.f64(
double %sub, double %b,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
%ret = call double @llvm.experimental.constrained.fsub.f64(
double -0.000000e+00,
double %mul,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
ret double %ret
}
; Verify that FP operations are not performed speculatively when FP exceptions
; are not being ignored.
;
; double f4(int n, double a) {
; // Because a + 1 may overflow, this should not be simplified.
; if (n > 0)
; return a + 1.0;
; return a;
; }
;
Add ‘llvm.experimental.constrained.fma‘ Intrinsic. Differential Revision: http://reviews.llvm.org/D36335 llvm-svn: 311629
2017-08-24 12:18:24 +08:00
;
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f4(i32 %n, double %a) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f4:
; X87: # %bb.0: # %entry
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: cmpl $0, {{[0-9]+}}(%esp)
; X87-NEXT: jle .LBB3_2
; X87-NEXT: # %bb.1: # %if.then
; X87-NEXT: fld1
; X87-NEXT: faddp %st, %st(1)
; X87-NEXT: .LBB3_2: # %if.end
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f4:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: cmpl $0, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: jle .LBB3_2
; X86-SSE-NEXT: # %bb.1: # %if.then
; X86-SSE-NEXT: addsd {{\.LCPI.*}}, %xmm0
; X86-SSE-NEXT: .LBB3_2: # %if.end
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: fldl (%esp)
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f4:
; SSE: # %bb.0: # %entry
; SSE-NEXT: testl %edi, %edi
; SSE-NEXT: jle .LBB3_2
; SSE-NEXT: # %bb.1: # %if.then
; SSE-NEXT: addsd {{.*}}(%rip), %xmm0
; SSE-NEXT: .LBB3_2: # %if.end
; SSE-NEXT: retq
;
; AVX-LABEL: f4:
; AVX: # %bb.0: # %entry
; AVX-NEXT: testl %edi, %edi
; AVX-NEXT: jle .LBB3_2
; AVX-NEXT: # %bb.1: # %if.then
; AVX-NEXT: vaddsd {{.*}}(%rip), %xmm0, %xmm0
; AVX-NEXT: .LBB3_2: # %if.end
; AVX-NEXT: retq
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
entry:
%cmp = icmp sgt i32 %n, 0
br i1 %cmp, label %if.then, label %if.end
if.then:
%add = call double @llvm.experimental.constrained.fadd.f64(
double 1.000000e+00, double %a,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
br label %if.end
if.end:
%a.0 = phi double [%add, %if.then], [ %a, %entry ]
ret double %a.0
}
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
; Verify that sqrt(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f5() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f5:
; X87: # %bb.0: # %entry
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fsqrt
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f5:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: sqrtsd %xmm0, %xmm0
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: fldl (%esp)
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f5:
; SSE: # %bb.0: # %entry
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: sqrtsd %xmm0, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: f5:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vsqrtsd %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.sqrt.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that pow(42.1, 3.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f6() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f6:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $28, %esp
; X87-NEXT: .cfi_def_cfa_offset 32
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl {{[0-9]+}}(%esp)
; X87-NEXT: fldl {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll pow
; X87-NEXT: addl $28, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f6:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $28, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 32
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll pow
; X86-SSE-NEXT: addl $28, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f6:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq pow
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f6:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq pow
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.pow.f64(double 42.1,
double 3.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that powi(42.1, 3) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f7() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f7:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: movl $3, {{[0-9]+}}(%esp)
; X87-NEXT: calll __powidf2
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f7:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: movl $3, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: calll __powidf2
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f7:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: movl $3, %edi
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq __powidf2
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f7:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: movl $3, %edi
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq __powidf2
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.powi.f64(double 42.1,
i32 3,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that sin(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f8() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f8:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll sin
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f8:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll sin
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f8:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq sin
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f8:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq sin
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.sin.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that cos(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f9() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f9:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll cos
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f9:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll cos
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f9:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq cos
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f9:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq cos
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.cos.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that exp(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f10() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f10:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll exp
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f10:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll exp
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f10:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq exp
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f10:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq exp
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.exp.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that exp2(42.1) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f11() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f11:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll exp2
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f11:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll exp2
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f11:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq exp2
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f11:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq exp2
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.exp2.f64(double 42.1,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that log(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f12() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f12:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll log
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f12:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll log
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f12:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq log
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f12:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq log
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.log.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that log10(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f13() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f13:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll log10
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f13:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll log10
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f13:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq log10
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f13:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq log10
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.log10.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that log2(42.0) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f14() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f14:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll log2
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f14:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll log2
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f14:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq log2
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f14:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq log2
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.log2.f64(double 42.0,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that rint(42.1) isn't simplified when the rounding mode is unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f15() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f15:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll rint
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f15:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll rint
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f15:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq rint
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f15:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vroundsd $4, %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.rint.f64(double 42.1,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
; Verify that nearbyint(42.1) isn't simplified when the rounding mode is
; unknown.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f16() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f16:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{\.LCPI.*}}
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll nearbyint
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f16:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll nearbyint
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f16:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq nearbyint
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f16:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vroundsd $12, %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
entry:
%result = call double @llvm.experimental.constrained.nearbyint.f64(
double 42.1,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
ret double %result
}
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f19() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f19:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $28, %esp
; X87-NEXT: .cfi_def_cfa_offset 32
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fstpl {{[0-9]+}}(%esp)
; X87-NEXT: movl $1072693248, {{[0-9]+}}(%esp) # imm = 0x3FF00000
; X87-NEXT: movl $0, (%esp)
; X87-NEXT: calll fmod
; X87-NEXT: addl $28, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f19:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $28, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 32
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll fmod
; X86-SSE-NEXT: addl $28, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f19:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: callq fmod
; SSE-NEXT: popq %rax
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
;
; AVX-LABEL: f19:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: callq fmod
; AVX-NEXT: popq %rax
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Support constrained FREM intrinsic Differential Revision: https://reviews.llvm.org/D50975 llvm-svn: 340201
2018-08-21 03:28:56 +08:00
entry:
%rem = call double @llvm.experimental.constrained.frem.f64(
double 1.000000e+00,
double 1.000000e+01,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
[FPEnv] Support constrained FREM intrinsic Differential Revision: https://reviews.llvm.org/D50975 llvm-svn: 340201
2018-08-21 03:28:56 +08:00
ret double %rem
}
[LegalizeDAG][X86] Enable STRICT_FP_TO_SINT/UINT to be promoted Differential Revision: https://reviews.llvm.org/D70220
2019-11-20 08:07:50 +08:00
; Verify that fptosi(%x) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
; FIXME: The SSE/AVX code does not raise an invalid exception for all values
; that don't fit in i8.
define i8 @f20s8(double %x) #0 {
; X87-LABEL: f20s8:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 12
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistps {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: movb {{[0-9]+}}(%esp), %al
; X87-NEXT: addl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
; X86-SSE-LABEL: f20s8:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: cvttsd2si {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f20s8:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvttsd2si %xmm0, %eax
; SSE-NEXT: # kill: def $al killed $al killed $eax
; SSE-NEXT: retq
;
; AVX-LABEL: f20s8:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvttsd2si %xmm0, %eax
; AVX-NEXT: # kill: def $al killed $al killed $eax
; AVX-NEXT: retq
entry:
%result = call i8 @llvm.experimental.constrained.fptosi.i8.f64(double %x,
metadata !"fpexcept.strict") #0
ret i8 %result
}
[X86] Add fptosi test to fp-intrinsics.ll
2019-11-12 15:30:55 +08:00
; Verify that fptosi(%x) isn't simplified when the rounding mode is
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; unknown.
; Verify that no gross errors happen.
; FIXME: The SSE/AVX code does not raise an invalid exception for all values
; that don't fit in i16.
define i16 @f20s16(double %x) #0 {
; X87-LABEL: f20s16:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 12
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistps {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: addl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
; X86-SSE-LABEL: f20s16:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: cvttsd2si {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: # kill: def $ax killed $ax killed $eax
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f20s16:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvttsd2si %xmm0, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
; SSE-NEXT: retq
;
; AVX-LABEL: f20s16:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvttsd2si %xmm0, %eax
; AVX-NEXT: # kill: def $ax killed $ax killed $eax
; AVX-NEXT: retq
entry:
%result = call i16 @llvm.experimental.constrained.fptosi.i16.f64(double %x,
metadata !"fpexcept.strict") #0
ret i16 %result
}
; Verify that fptosi(%x) isn't simplified when the rounding mode is
; unknown.
[X86] Add fptosi test to fp-intrinsics.ll
2019-11-12 15:30:55 +08:00
; Verify that no gross errors happen.
define i32 @f20s(double %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f20s:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 12
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw (%esp)
; X87-NEXT: movzwl (%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistpl {{[0-9]+}}(%esp)
; X87-NEXT: fldcw (%esp)
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: addl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f20s:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: cvttsd2si {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: retl
;
[X86] Add fptosi test to fp-intrinsics.ll
2019-11-12 15:30:55 +08:00
; SSE-LABEL: f20s:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvttsd2si %xmm0, %eax
; SSE-NEXT: retq
;
; AVX-LABEL: f20s:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvttsd2si %xmm0, %eax
; AVX-NEXT: retq
entry:
%result = call i32 @llvm.experimental.constrained.fptosi.i32.f64(double %x,
metadata !"fpexcept.strict") #0
ret i32 %result
}
[FPEnv] Add fptosi and fptoui constrained intrinsics. This implements constrained floating point intrinsics for FP to signed and unsigned integers. Quoting from D32319: The purpose of the constrained intrinsics is to force the optimizer to respect the restrictions that will be necessary to support things like the STDC FENV_ACCESS ON pragma without interfering with optimizations when these restrictions are not needed. Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton Approved by: Craig Topper Differential Revision: http://reviews.llvm.org/D63782 llvm-svn: 370228
2019-08-29 00:33:36 +08:00
; Verify that fptoui(%x) isn't simplified when the rounding mode is
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; unknown.
; Verify that no gross errors happen.
; FIXME: This code generates spurious inexact exceptions.
define i64 @f20s64(double %x) #0 {
; X87-LABEL: f20s64:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $20, %esp
; X87-NEXT: .cfi_def_cfa_offset 24
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistpll {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: movl {{[0-9]+}}(%esp), %edx
; X87-NEXT: addl $20, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
; X86-SSE-LABEL: f20s64:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $20, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 24
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fldl {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fnstcw {{[0-9]+}}(%esp)
; X86-SSE-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: orl $3072, %eax # imm = 0xC00
; X86-SSE-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fldcw {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fistpll {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fldcw {{[0-9]+}}(%esp)
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %edx
; X86-SSE-NEXT: addl $20, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f20s64:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvttsd2si %xmm0, %rax
; SSE-NEXT: retq
;
; AVX-LABEL: f20s64:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvttsd2si %xmm0, %rax
; AVX-NEXT: retq
entry:
%result = call i64 @llvm.experimental.constrained.fptosi.i64.f64(double %x,
metadata !"fpexcept.strict") #0
ret i64 %result
}
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; Verify that fptoui(%x) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
define i128 @f20s128(double %x) nounwind strictfp {
; X87-LABEL: f20s128:
; X87: # %bb.0: # %entry
; X87-NEXT: pushl %edi
; X87-NEXT: pushl %esi
; X87-NEXT: subl $36, %esp
; X87-NEXT: movl {{[0-9]+}}(%esp), %esi
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fstpl {{[0-9]+}}(%esp)
; X87-NEXT: leal {{[0-9]+}}(%esp), %eax
; X87-NEXT: movl %eax, (%esp)
; X87-NEXT: calll __fixdfti
; X87-NEXT: subl $4, %esp
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X87-NEXT: movl {{[0-9]+}}(%esp), %edx
; X87-NEXT: movl {{[0-9]+}}(%esp), %edi
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X87-NEXT: movl %edi, 12(%esi)
; X87-NEXT: movl %edx, 8(%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X87-NEXT: movl %ecx, 4(%esi)
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X87-NEXT: movl %eax, (%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X87-NEXT: movl %esi, %eax
; X87-NEXT: addl $36, %esp
; X87-NEXT: popl %esi
; X87-NEXT: popl %edi
; X87-NEXT: retl $4
;
; X86-SSE-LABEL: f20s128:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: pushl %edi
; X86-SSE-NEXT: pushl %esi
; X86-SSE-NEXT: subl $36, %esp
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %esi
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: leal {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: movl %eax, (%esp)
; X86-SSE-NEXT: calll __fixdfti
; X86-SSE-NEXT: subl $4, %esp
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %edx
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %edi
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X86-SSE-NEXT: movl %edi, 12(%esi)
; X86-SSE-NEXT: movl %edx, 8(%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X86-SSE-NEXT: movl %ecx, 4(%esi)
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X86-SSE-NEXT: movl %eax, (%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X86-SSE-NEXT: movl %esi, %eax
; X86-SSE-NEXT: addl $36, %esp
; X86-SSE-NEXT: popl %esi
; X86-SSE-NEXT: popl %edi
; X86-SSE-NEXT: retl $4
;
; SSE-LABEL: f20s128:
; SSE: # %bb.0: # %entry
; SSE-NEXT: pushq %rax
; SSE-NEXT: callq __fixdfti
; SSE-NEXT: popq %rcx
; SSE-NEXT: retq
;
; AVX-LABEL: f20s128:
; AVX: # %bb.0: # %entry
; AVX-NEXT: pushq %rax
; AVX-NEXT: callq __fixdfti
; AVX-NEXT: popq %rcx
; AVX-NEXT: retq
entry:
%result = call i128 @llvm.experimental.constrained.fptosi.i128.f64(double %x,
metadata !"fpexcept.strict") #0
ret i128 %result
}
[LegalizeDAG][X86] Enable STRICT_FP_TO_SINT/UINT to be promoted Differential Revision: https://reviews.llvm.org/D70220
2019-11-20 08:07:50 +08:00
; Verify that fptoui(%x) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
; FIXME: The SSE/AVX code does not raise an invalid exception for all values
; that don't fit in i8.
define i8 @f20u8(double %x) #0 {
; X87-LABEL: f20u8:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 12
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistps {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: movb {{[0-9]+}}(%esp), %al
; X87-NEXT: addl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
; X86-SSE-LABEL: f20u8:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: cvttsd2si {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: # kill: def $al killed $al killed $eax
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f20u8:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvttsd2si %xmm0, %eax
; SSE-NEXT: # kill: def $al killed $al killed $eax
; SSE-NEXT: retq
;
; AVX-LABEL: f20u8:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvttsd2si %xmm0, %eax
; AVX-NEXT: # kill: def $al killed $al killed $eax
; AVX-NEXT: retq
entry:
%result = call i8 @llvm.experimental.constrained.fptoui.i8.f64(double %x,
metadata !"fpexcept.strict") #0
ret i8 %result
}
; Verify that fptoui(%x) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
; FIXME: The SSE/AVX code does not raise an invalid exception for all values
; that don't fit in i16.
define i16 @f20u16(double %x) #0 {
; X87-LABEL: f20u16:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 12
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw (%esp)
; X87-NEXT: movzwl (%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistpl {{[0-9]+}}(%esp)
; X87-NEXT: fldcw (%esp)
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: # kill: def $ax killed $ax killed $eax
; X87-NEXT: addl $8, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
; X86-SSE-LABEL: f20u16:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: cvttsd2si {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: # kill: def $ax killed $ax killed $eax
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f20u16:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvttsd2si %xmm0, %eax
; SSE-NEXT: # kill: def $ax killed $ax killed $eax
; SSE-NEXT: retq
;
; AVX-LABEL: f20u16:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvttsd2si %xmm0, %eax
; AVX-NEXT: # kill: def $ax killed $ax killed $eax
; AVX-NEXT: retq
entry:
%result = call i16 @llvm.experimental.constrained.fptoui.i16.f64(double %x,
metadata !"fpexcept.strict") #0
ret i16 %result
}
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; Verify that fptoui(%x) isn't simplified when the rounding mode is
; unknown.
[FPEnv] Add fptosi and fptoui constrained intrinsics. This implements constrained floating point intrinsics for FP to signed and unsigned integers. Quoting from D32319: The purpose of the constrained intrinsics is to force the optimizer to respect the restrictions that will be necessary to support things like the STDC FENV_ACCESS ON pragma without interfering with optimizations when these restrictions are not needed. Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton Approved by: Craig Topper Differential Revision: http://reviews.llvm.org/D63782 llvm-svn: 370228
2019-08-29 00:33:36 +08:00
; Verify that no gross errors happen.
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; FIXME: The X87/SSE/AVX1 code does not raise an invalid exception for all
; values that don't fit in i32. The AVX512 code does.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define i32 @f20u(double %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f20u:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $20, %esp
; X87-NEXT: .cfi_def_cfa_offset 24
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fnstcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistpll {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: addl $20, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f20u:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
; X86-SSE-NEXT: movapd %xmm0, %xmm2
; X86-SSE-NEXT: cmpltsd %xmm1, %xmm2
; X86-SSE-NEXT: movapd %xmm2, %xmm3
; X86-SSE-NEXT: andpd %xmm0, %xmm2
; X86-SSE-NEXT: xorl %eax, %eax
; X86-SSE-NEXT: ucomisd %xmm0, %xmm1
; X86-SSE-NEXT: subsd %xmm1, %xmm0
; X86-SSE-NEXT: andnpd %xmm0, %xmm3
; X86-SSE-NEXT: orpd %xmm3, %xmm2
; X86-SSE-NEXT: cvttsd2si %xmm2, %ecx
; X86-SSE-NEXT: setbe %al
; X86-SSE-NEXT: shll $31, %eax
; X86-SSE-NEXT: xorl %ecx, %eax
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f20u:
; SSE: # %bb.0: # %entry
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; SSE-NEXT: cvttsd2si %xmm0, %rax
; SSE-NEXT: # kill: def $eax killed $eax killed $rax
; SSE-NEXT: retq
;
; AVX1-LABEL: f20u:
; AVX1: # %bb.0: # %entry
; AVX1-NEXT: vcvttsd2si %xmm0, %rax
; AVX1-NEXT: # kill: def $eax killed $eax killed $rax
; AVX1-NEXT: retq
;
; AVX512-LABEL: f20u:
; AVX512: # %bb.0: # %entry
; AVX512-NEXT: vcvttsd2usi %xmm0, %eax
; AVX512-NEXT: retq
entry:
%result = call i32 @llvm.experimental.constrained.fptoui.i32.f64(double %x,
metadata !"fpexcept.strict") #0
ret i32 %result
}
; Verify that fptoui(%x) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
; FIXME: This code generates spurious inexact exceptions.
define i64 @f20u64(double %x) #0 {
; X87-LABEL: f20u64:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $20, %esp
; X87-NEXT: .cfi_def_cfa_offset 24
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: flds {{\.LCPI.*}}
; X87-NEXT: fld %st(1)
; X87-NEXT: fsub %st(1), %st
; X87-NEXT: xorl %edx, %edx
; X87-NEXT: fxch %st(1)
; X87-NEXT: fucompi %st(2), %st
; X87-NEXT: fcmovnbe %st(1), %st
; X87-NEXT: fstp %st(1)
; X87-NEXT: setbe %dl
; X87-NEXT: fnstcw {{[0-9]+}}(%esp)
; X87-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X87-NEXT: orl $3072, %eax # imm = 0xC00
; X87-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: fistpll {{[0-9]+}}(%esp)
; X87-NEXT: fldcw {{[0-9]+}}(%esp)
; X87-NEXT: shll $31, %edx
; X87-NEXT: xorl {{[0-9]+}}(%esp), %edx
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: addl $20, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
; X86-SSE-LABEL: f20u64:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $20, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 24
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
; X86-SSE-NEXT: movapd %xmm0, %xmm2
; X86-SSE-NEXT: subsd %xmm1, %xmm2
; X86-SSE-NEXT: movapd %xmm0, %xmm3
; X86-SSE-NEXT: cmpltsd %xmm1, %xmm3
; X86-SSE-NEXT: movapd %xmm3, %xmm4
; X86-SSE-NEXT: andnpd %xmm2, %xmm4
; X86-SSE-NEXT: andpd %xmm0, %xmm3
; X86-SSE-NEXT: orpd %xmm4, %xmm3
; X86-SSE-NEXT: movlpd %xmm3, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fldl {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fnstcw {{[0-9]+}}(%esp)
; X86-SSE-NEXT: movzwl {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: orl $3072, %eax # imm = 0xC00
; X86-SSE-NEXT: movw %ax, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fldcw {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fistpll {{[0-9]+}}(%esp)
; X86-SSE-NEXT: fldcw {{[0-9]+}}(%esp)
; X86-SSE-NEXT: xorl %edx, %edx
; X86-SSE-NEXT: ucomisd %xmm0, %xmm1
; X86-SSE-NEXT: setbe %dl
; X86-SSE-NEXT: shll $31, %edx
; X86-SSE-NEXT: xorl {{[0-9]+}}(%esp), %edx
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: addl $20, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f20u64:
; SSE: # %bb.0: # %entry
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
; SSE-NEXT: movapd %xmm0, %xmm2
; SSE-NEXT: cmpltsd %xmm1, %xmm2
; SSE-NEXT: movapd %xmm2, %xmm3
; SSE-NEXT: andpd %xmm0, %xmm2
; SSE-NEXT: xorl %eax, %eax
; SSE-NEXT: ucomisd %xmm1, %xmm0
; SSE-NEXT: subsd %xmm1, %xmm0
; SSE-NEXT: andnpd %xmm0, %xmm3
; SSE-NEXT: orpd %xmm3, %xmm2
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; SSE-NEXT: cvttsd2si %xmm2, %rcx
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: setae %al
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; SSE-NEXT: shlq $63, %rax
; SSE-NEXT: xorq %rcx, %rax
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-NEXT: retq
;
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; AVX1-LABEL: f20u64:
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX1: # %bb.0: # %entry
; AVX1-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
; AVX1-NEXT: vcmpltsd %xmm1, %xmm0, %xmm2
; AVX1-NEXT: vsubsd %xmm1, %xmm0, %xmm3
; AVX1-NEXT: vblendvpd %xmm2, %xmm0, %xmm3, %xmm2
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; AVX1-NEXT: vcvttsd2si %xmm2, %rcx
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX1-NEXT: xorl %eax, %eax
; AVX1-NEXT: vucomisd %xmm1, %xmm0
; AVX1-NEXT: setae %al
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; AVX1-NEXT: shlq $63, %rax
; AVX1-NEXT: xorq %rcx, %rax
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX1-NEXT: retq
;
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; AVX512-LABEL: f20u64:
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX512: # %bb.0: # %entry
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; AVX512-NEXT: vcvttsd2usi %xmm0, %rax
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; AVX512-NEXT: retq
[FPEnv] Add fptosi and fptoui constrained intrinsics. This implements constrained floating point intrinsics for FP to signed and unsigned integers. Quoting from D32319: The purpose of the constrained intrinsics is to force the optimizer to respect the restrictions that will be necessary to support things like the STDC FENV_ACCESS ON pragma without interfering with optimizations when these restrictions are not needed. Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton Approved by: Craig Topper Differential Revision: http://reviews.llvm.org/D63782 llvm-svn: 370228
2019-08-29 00:33:36 +08:00
entry:
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
%result = call i64 @llvm.experimental.constrained.fptoui.i64.f64(double %x,
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
ret i64 %result
[FPEnv] Add fptosi and fptoui constrained intrinsics. This implements constrained floating point intrinsics for FP to signed and unsigned integers. Quoting from D32319: The purpose of the constrained intrinsics is to force the optimizer to respect the restrictions that will be necessary to support things like the STDC FENV_ACCESS ON pragma without interfering with optimizations when these restrictions are not needed. Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton Approved by: Craig Topper Differential Revision: http://reviews.llvm.org/D63782 llvm-svn: 370228
2019-08-29 00:33:36 +08:00
}
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; Verify that fptoui(%x) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
define i128 @f20u128(double %x) nounwind strictfp {
; X87-LABEL: f20u128:
; X87: # %bb.0: # %entry
; X87-NEXT: pushl %edi
; X87-NEXT: pushl %esi
; X87-NEXT: subl $36, %esp
; X87-NEXT: movl {{[0-9]+}}(%esp), %esi
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fstpl {{[0-9]+}}(%esp)
; X87-NEXT: leal {{[0-9]+}}(%esp), %eax
; X87-NEXT: movl %eax, (%esp)
; X87-NEXT: calll __fixunsdfti
; X87-NEXT: subl $4, %esp
; X87-NEXT: movl {{[0-9]+}}(%esp), %eax
; X87-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X87-NEXT: movl {{[0-9]+}}(%esp), %edx
; X87-NEXT: movl {{[0-9]+}}(%esp), %edi
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X87-NEXT: movl %edi, 12(%esi)
; X87-NEXT: movl %edx, 8(%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X87-NEXT: movl %ecx, 4(%esi)
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X87-NEXT: movl %eax, (%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X87-NEXT: movl %esi, %eax
; X87-NEXT: addl $36, %esp
; X87-NEXT: popl %esi
; X87-NEXT: popl %edi
; X87-NEXT: retl $4
;
; X86-SSE-LABEL: f20u128:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: pushl %edi
; X86-SSE-NEXT: pushl %esi
; X86-SSE-NEXT: subl $36, %esp
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %esi
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, {{[0-9]+}}(%esp)
; X86-SSE-NEXT: leal {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: movl %eax, (%esp)
; X86-SSE-NEXT: calll __fixunsdfti
; X86-SSE-NEXT: subl $4, %esp
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %eax
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %ecx
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %edx
; X86-SSE-NEXT: movl {{[0-9]+}}(%esp), %edi
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X86-SSE-NEXT: movl %edi, 12(%esi)
; X86-SSE-NEXT: movl %edx, 8(%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X86-SSE-NEXT: movl %ecx, 4(%esi)
[FPEnv][SelectionDAG] Relax chain requirements This patch implements the following changes: 1) SelectionDAGBuilder::visitConstrainedFPIntrinsic currently treats each constrained intrinsic like a global barrier (e.g. a function call) and fully serializes all pending chains. This is actually not required; it is allowed for constrained intrinsics to be reordered w.r.t one another or (nonvolatile) memory accesses. The MI-level scheduler already allows for that flexibility, so it makes sense to allow it at the DAG level as well. This patch therefore changes the way chains for constrained intrisincs are created, and handles them basically like load operations are handled. This has the effect that constrained intrinsics are no longer serialized against one another or (nonvolatile) loads. They are still serialized against stores, but that seems hard to change with the current DAG chain setup, and it also doesn't seem to be a big problem preventing DAG 2) The OPC_CheckFoldableChainNode check requires that each of the intermediate nodes in a multi-node pattern match only has a single use. This check tends to fail if those intermediate nodes are strict operations as those have a chain output that typically indeed has another use. However, we don't really need to consider chains here at all, since they will all be rewritten anyway by UpdateChains later. Other parts of the matcher therefore already ignore chains, but this hasOneUse check doesn't. This patch replaces hasOneUse by a custom test that verifies there is no more than one use of any non-chain output value. In theory, this change could affect code unrelated to strict FP nodes, but at least on SystemZ I could not find any single instance of that happening 3) The SystemZ back-end currently does not allow matching multiply-and- extend operations (32x32 -> 64bit or 64x64 -> 128bit FP multiply) for strict FP operations. This was not possible in the past due to the problems described under 1) and 2) above. With those issues fixed, it is now possible to fully support those instructions in strict mode as well, and this patch does so. Differential Revision: https://reviews.llvm.org/D70913
2019-12-06 18:02:11 +08:00
; X86-SSE-NEXT: movl %eax, (%esi)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
; X86-SSE-NEXT: movl %esi, %eax
; X86-SSE-NEXT: addl $36, %esp
; X86-SSE-NEXT: popl %esi
; X86-SSE-NEXT: popl %edi
; X86-SSE-NEXT: retl $4
;
; SSE-LABEL: f20u128:
; SSE: # %bb.0: # %entry
; SSE-NEXT: pushq %rax
; SSE-NEXT: callq __fixunsdfti
; SSE-NEXT: popq %rcx
; SSE-NEXT: retq
;
; AVX-LABEL: f20u128:
; AVX: # %bb.0: # %entry
; AVX-NEXT: pushq %rax
; AVX-NEXT: callq __fixunsdfti
; AVX-NEXT: popq %rcx
; AVX-NEXT: retq
entry:
%result = call i128 @llvm.experimental.constrained.fptoui.i128.f64(double %x,
metadata !"fpexcept.strict") #0
ret i128 %result
}
Add constrained fptrunc and fpext intrinsics. The new fptrunc and fpext intrinsics are constrained versions of the regular fptrunc and fpext instructions. Reviewed by: Andrew Kaylor, Craig Topper, Cameron McInally, Conner Abbot Approved by: Craig Topper Differential Revision: https://reviews.llvm.org/D55897 llvm-svn: 360581
2019-05-13 21:23:30 +08:00
; Verify that round(42.1) isn't simplified when the rounding mode is
; unknown.
; Verify that no gross errors happen.
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define float @f21() #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f21:
; X87: # %bb.0: # %entry
; X87-NEXT: pushl %eax
; X87-NEXT: .cfi_def_cfa_offset 8
; X87-NEXT: fldl {{\.LCPI.*}}
; X87-NEXT: fstps (%esp)
; X87-NEXT: flds (%esp)
; X87-NEXT: popl %eax
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f21:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: pushl %eax
; X86-SSE-NEXT: .cfi_def_cfa_offset 8
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: cvtsd2ss %xmm0, %xmm0
; X86-SSE-NEXT: movss %xmm0, (%esp)
; X86-SSE-NEXT: flds (%esp)
; X86-SSE-NEXT: popl %eax
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f21:
; SSE: # %bb.0: # %entry
; SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; SSE-NEXT: cvtsd2ss %xmm0, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: f21:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
; AVX-NEXT: vcvtsd2ss %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
Add constrained fptrunc and fpext intrinsics. The new fptrunc and fpext intrinsics are constrained versions of the regular fptrunc and fpext instructions. Reviewed by: Andrew Kaylor, Craig Topper, Cameron McInally, Conner Abbot Approved by: Craig Topper Differential Revision: https://reviews.llvm.org/D55897 llvm-svn: 360581
2019-05-13 21:23:30 +08:00
entry:
%result = call float @llvm.experimental.constrained.fptrunc.f32.f64(
double 42.1,
metadata !"round.dynamic",
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained fptrunc and fpext intrinsics. The new fptrunc and fpext intrinsics are constrained versions of the regular fptrunc and fpext instructions. Reviewed by: Andrew Kaylor, Craig Topper, Cameron McInally, Conner Abbot Approved by: Craig Topper Differential Revision: https://reviews.llvm.org/D55897 llvm-svn: 360581
2019-05-13 21:23:30 +08:00
ret float %result
}
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
define double @f22(float %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f22:
; X87: # %bb.0: # %entry
; X87-NEXT: flds {{[0-9]+}}(%esp)
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f22:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE-NEXT: cvtss2sd %xmm0, %xmm0
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: fldl (%esp)
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
[X86] Autogenerate checks in fp-intrinsics.ll. Split fma tests off to fp-intrinsics-fma.ll. NFC
2019-11-12 07:42:25 +08:00
; SSE-LABEL: f22:
; SSE: # %bb.0: # %entry
; SSE-NEXT: cvtss2sd %xmm0, %xmm0
; SSE-NEXT: retq
;
; AVX-LABEL: f22:
; AVX: # %bb.0: # %entry
; AVX-NEXT: vcvtss2sd %xmm0, %xmm0, %xmm0
; AVX-NEXT: retq
Add constrained fptrunc and fpext intrinsics. The new fptrunc and fpext intrinsics are constrained versions of the regular fptrunc and fpext instructions. Reviewed by: Andrew Kaylor, Craig Topper, Cameron McInally, Conner Abbot Approved by: Craig Topper Differential Revision: https://reviews.llvm.org/D55897 llvm-svn: 360581
2019-05-13 21:23:30 +08:00
entry:
%result = call double @llvm.experimental.constrained.fpext.f64.f32(float %x,
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
metadata !"fpexcept.strict") #0
Add constrained fptrunc and fpext intrinsics. The new fptrunc and fpext intrinsics are constrained versions of the regular fptrunc and fpext instructions. Reviewed by: Andrew Kaylor, Craig Topper, Cameron McInally, Conner Abbot Approved by: Craig Topper Differential Revision: https://reviews.llvm.org/D55897 llvm-svn: 360581
2019-05-13 21:23:30 +08:00
ret double %result
}
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
define i32 @f23(double %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f23:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll lrint
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; X86-SSE-LABEL: f23:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll lrint
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f23:
; SSE: # %bb.0: # %entry
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq lrint
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
;
; AVX-LABEL: f23:
; AVX: # %bb.0: # %entry
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq lrint
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i32 @llvm.experimental.constrained.lrint.i32.f64(double %x,
metadata !"round.dynamic",
metadata !"fpexcept.strict") #0
ret i32 %result
}
define i32 @f24(float %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f24:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{[0-9]+}}(%esp)
; X87-NEXT: fstps (%esp)
; X87-NEXT: calll lrintf
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; X86-SSE-LABEL: f24:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE-NEXT: movss %xmm0, (%esp)
; X86-SSE-NEXT: calll lrintf
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f24:
; SSE: # %bb.0: # %entry
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq lrintf
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
;
; AVX-LABEL: f24:
; AVX: # %bb.0: # %entry
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq lrintf
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i32 @llvm.experimental.constrained.lrint.i32.f32(float %x,
metadata !"round.dynamic",
metadata !"fpexcept.strict") #0
ret i32 %result
}
define i64 @f25(double %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f25:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll llrint
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f25:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll llrint
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f25:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq llrint
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
;
; AVX-LABEL: f25:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq llrint
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i64 @llvm.experimental.constrained.llrint.i64.f64(double %x,
metadata !"round.dynamic",
metadata !"fpexcept.strict") #0
ret i64 %result
}
define i64 @f26(float %x) {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f26:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{[0-9]+}}(%esp)
; X87-NEXT: fstps (%esp)
; X87-NEXT: calll llrintf
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f26:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE-NEXT: movss %xmm0, (%esp)
; X86-SSE-NEXT: calll llrintf
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f26:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq llrintf
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
;
; AVX-LABEL: f26:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq llrintf
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i64 @llvm.experimental.constrained.llrint.i64.f32(float %x,
metadata !"round.dynamic",
metadata !"fpexcept.strict") #0
ret i64 %result
}
define i32 @f27(double %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f27:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll lround
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; X86-SSE-LABEL: f27:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll lround
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f27:
; SSE: # %bb.0: # %entry
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq lround
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
;
; AVX-LABEL: f27:
; AVX: # %bb.0: # %entry
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq lround
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i32 @llvm.experimental.constrained.lround.i32.f64(double %x,
metadata !"fpexcept.strict") #0
ret i32 %result
}
define i32 @f28(float %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f28:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{[0-9]+}}(%esp)
; X87-NEXT: fstps (%esp)
; X87-NEXT: calll lroundf
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; X86-SSE-LABEL: f28:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE-NEXT: movss %xmm0, (%esp)
; X86-SSE-NEXT: calll lroundf
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f28:
; SSE: # %bb.0: # %entry
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq lroundf
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
;
; AVX-LABEL: f28:
; AVX: # %bb.0: # %entry
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq lroundf
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i32 @llvm.experimental.constrained.lround.i32.f32(float %x,
metadata !"fpexcept.strict") #0
ret i32 %result
}
define i64 @f29(double %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f29:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: fldl {{[0-9]+}}(%esp)
; X87-NEXT: fstpl (%esp)
; X87-NEXT: calll llround
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f29:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
; X86-SSE-NEXT: movsd %xmm0, (%esp)
; X86-SSE-NEXT: calll llround
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f29:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq llround
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
;
; AVX-LABEL: f29:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq llround
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i64 @llvm.experimental.constrained.llround.i64.f64(double %x,
metadata !"fpexcept.strict") #0
ret i64 %result
}
define i64 @f30(float %x) #0 {
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
; X87-LABEL: f30:
; X87: # %bb.0: # %entry
; X87-NEXT: subl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 16
; X87-NEXT: flds {{[0-9]+}}(%esp)
; X87-NEXT: fstps (%esp)
; X87-NEXT: calll llroundf
; X87-NEXT: addl $12, %esp
; X87-NEXT: .cfi_def_cfa_offset 4
; X87-NEXT: retl
;
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
; X86-SSE-LABEL: f30:
; X86-SSE: # %bb.0: # %entry
; X86-SSE-NEXT: subl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 16
; X86-SSE-NEXT: movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
; X86-SSE-NEXT: movss %xmm0, (%esp)
; X86-SSE-NEXT: calll llroundf
; X86-SSE-NEXT: addl $12, %esp
; X86-SSE-NEXT: .cfi_def_cfa_offset 4
; X86-SSE-NEXT: retl
;
; SSE-LABEL: f30:
; SSE: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; SSE-NEXT: pushq %rax
; SSE-NEXT: .cfi_def_cfa_offset 16
; SSE-NEXT: callq llroundf
; SSE-NEXT: popq %rcx
; SSE-NEXT: .cfi_def_cfa_offset 8
; SSE-NEXT: retq
[LegalizeTypes][X86] Add support for expanding the result type of STRICT_LLROUND and STRICT_LLRINT. This doesn't handle softening the input type, but we don't handle softening any of the strict nodes yet. Skipping that made it easy to reuse an existing function for creating a libcall from a node with a chain.
2019-11-18 11:58:11 +08:00
;
; AVX-LABEL: f30:
; AVX: # %bb.0: # %entry
[LegalizeDAG] Convert strict fp nodes to libcalls without losing the chain. Previously we mutated the node and then converted it to a libcall. But this loses the chain information. This patch keeps the chain, but unfortunately breaks tail call optimization as the functions involved in deciding if a node is in tail call position can't handle the chain. But correct ordering seems more important to be right. Somehow the SystemZ tests improved. I looked at one of them and it seemed that we're handling the split vector elements in a different order and that made the copies work better. Differential Revision: https://reviews.llvm.org/D70334
2019-11-19 03:14:13 +08:00
; AVX-NEXT: pushq %rax
; AVX-NEXT: .cfi_def_cfa_offset 16
; AVX-NEXT: callq llroundf
; AVX-NEXT: popq %rcx
; AVX-NEXT: .cfi_def_cfa_offset 8
; AVX-NEXT: retq
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
entry:
%result = call i64 @llvm.experimental.constrained.llround.i64.f32(float %x,
metadata !"fpexcept.strict") #0
ret i64 %result
}
[FPEnv] Strict FP tests should use the requisite function attributes. A set of function attributes is required in any function that uses constrained floating point intrinsics. None of our tests use these attributes. This patch fixes this. These tests have been tested against the IR verifier changes in D68233. Reviewed by: andrew.w.kaylor, cameron.mcinally, uweigand Approved by: andrew.w.kaylor Differential Revision: https://reviews.llvm.org/D67925 llvm-svn: 373761
2019-10-05 01:03:46 +08:00
attributes #0 = { strictfp }
Add intrinsics for constrained floating point operations This commit introduces a set of experimental intrinsics intended to prevent optimizations that make assumptions about the rounding mode and floating point exception behavior. These intrinsics will later be extended to specify flush-to-zero behavior. More work is also required to model instruction dependencies in machine code and to generate these instructions from clang (when required by pragmas and/or command line options that are not currently supported). Differential Revision: https://reviews.llvm.org/D27028 llvm-svn: 293226
2017-01-27 07:27:59 +08:00
@llvm.fp.env = thread_local global i8 zeroinitializer, section "llvm.metadata"
declare double @llvm.experimental.constrained.fadd.f64(double, double, metadata, metadata)
declare double @llvm.experimental.constrained.fsub.f64(double, double, metadata, metadata)
[FPEnv] Support constrained FREM intrinsic Differential Revision: https://reviews.llvm.org/D50975 llvm-svn: 340201
2018-08-21 03:28:56 +08:00
declare double @llvm.experimental.constrained.fmul.f64(double, double, metadata, metadata)
declare double @llvm.experimental.constrained.fdiv.f64(double, double, metadata, metadata)
declare double @llvm.experimental.constrained.frem.f64(double, double, metadata, metadata)
Add constrained intrinsics for some libm-equivalent operations Differential revision: https://reviews.llvm.org/D32319 llvm-svn: 303922
2017-05-26 05:31:00 +08:00
declare double @llvm.experimental.constrained.sqrt.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.pow.f64(double, double, metadata, metadata)
declare double @llvm.experimental.constrained.powi.f64(double, i32, metadata, metadata)
declare double @llvm.experimental.constrained.sin.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.cos.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.exp.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.exp2.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.log.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.log10.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.log2.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.rint.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.nearbyint.f64(double, metadata, metadata)
[LegalizeDAG][X86] Enable STRICT_FP_TO_SINT/UINT to be promoted Differential Revision: https://reviews.llvm.org/D70220
2019-11-20 08:07:50 +08:00
declare i8 @llvm.experimental.constrained.fptosi.i8.f64(double, metadata)
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
declare i16 @llvm.experimental.constrained.fptosi.i16.f64(double, metadata)
[FPEnv] Add fptosi and fptoui constrained intrinsics. This implements constrained floating point intrinsics for FP to signed and unsigned integers. Quoting from D32319: The purpose of the constrained intrinsics is to force the optimizer to respect the restrictions that will be necessary to support things like the STDC FENV_ACCESS ON pragma without interfering with optimizations when these restrictions are not needed. Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton Approved by: Craig Topper Differential Revision: http://reviews.llvm.org/D63782 llvm-svn: 370228
2019-08-29 00:33:36 +08:00
declare i32 @llvm.experimental.constrained.fptosi.i32.f64(double, metadata)
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
declare i64 @llvm.experimental.constrained.fptosi.i64.f64(double, metadata)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
declare i128 @llvm.experimental.constrained.fptosi.i128.f64(double, metadata)
[LegalizeDAG][X86] Enable STRICT_FP_TO_SINT/UINT to be promoted Differential Revision: https://reviews.llvm.org/D70220
2019-11-20 08:07:50 +08:00
declare i8 @llvm.experimental.constrained.fptoui.i8.f64(double, metadata)
declare i16 @llvm.experimental.constrained.fptoui.i16.f64(double, metadata)
[FPEnv] Add fptosi and fptoui constrained intrinsics. This implements constrained floating point intrinsics for FP to signed and unsigned integers. Quoting from D32319: The purpose of the constrained intrinsics is to force the optimizer to respect the restrictions that will be necessary to support things like the STDC FENV_ACCESS ON pragma without interfering with optimizations when these restrictions are not needed. Reviewed by: Andrew Kaylor, Craig Topper, Hal Finkel, Cameron McInally, Roman Lebedev, Kit Barton Approved by: Craig Topper Differential Revision: http://reviews.llvm.org/D63782 llvm-svn: 370228
2019-08-29 00:33:36 +08:00
declare i32 @llvm.experimental.constrained.fptoui.i32.f64(double, metadata)
[X86] Add custom type legalization and lowering for scalar STRICT_FP_TO_SINT/UINT This is a first pass at Custom lowering for these operations. I also updated some of the vector code where it was obviously easy and straightforward. More work needed in follow up. This enables these operations to be handled with X87 where special rounding control adjustments are needed to perform a truncate. Still need to fix Promotion in the target independent code in LegalizeDAG. llrint/llround split into separate test file because we can't make a strict libcall properly yet either and we need to do that when i64 isn't a legal type. This does not include any isel support. So we still rely on the mutation in SelectionDAGIsel to remove the strict from this stuff later. Except for the X87 stuff which goes through custom nodes that already had chains. Differential Revision: https://reviews.llvm.org/D70214
2019-11-20 07:51:19 +08:00
declare i64 @llvm.experimental.constrained.fptoui.i64.f64(double, metadata)
[LegalizeTypes][X86] Add ExpandIntegerResult support for STRICT_FP_TO_SINT/STRICT_FP_TO_UINT.
2019-11-28 09:44:43 +08:00
declare i128 @llvm.experimental.constrained.fptoui.i128.f64(double, metadata)
Add constrained fptrunc and fpext intrinsics. The new fptrunc and fpext intrinsics are constrained versions of the regular fptrunc and fpext instructions. Reviewed by: Andrew Kaylor, Craig Topper, Cameron McInally, Conner Abbot Approved by: Craig Topper Differential Revision: https://reviews.llvm.org/D55897 llvm-svn: 360581
2019-05-13 21:23:30 +08:00
declare float @llvm.experimental.constrained.fptrunc.f32.f64(double, metadata, metadata)
declare double @llvm.experimental.constrained.fpext.f64.f32(float, metadata)
[FPEnv] Add constrained intrinsics for lrint and lround Earlier in the year intrinsics for lrint, llrint, lround and llround were added to llvm. The constrained versions are now implemented here. Reviewed by: andrew.w.kaylor, craig.topper, cameron.mcinally Approved by: craig.topper Differential Revision: https://reviews.llvm.org/D64746 llvm-svn: 373900
2019-10-07 21:20:00 +08:00
declare i32 @llvm.experimental.constrained.lrint.i32.f64(double, metadata, metadata)
declare i32 @llvm.experimental.constrained.lrint.i32.f32(float, metadata, metadata)
declare i64 @llvm.experimental.constrained.llrint.i64.f64(double, metadata, metadata)
declare i64 @llvm.experimental.constrained.llrint.i64.f32(float, metadata, metadata)
declare i32 @llvm.experimental.constrained.lround.i32.f64(double, metadata)
declare i32 @llvm.experimental.constrained.lround.i32.f32(float, metadata)
declare i64 @llvm.experimental.constrained.llround.i64.f64(double, metadata)
declare i64 @llvm.experimental.constrained.llround.i64.f32(float, metadata)